xref: /openbmc/linux/drivers/infiniband/sw/rdmavt/cq.c (revision d164bf64)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2016 - 2018 Intel Corporation.
4  */
5 
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include "cq.h"
9 #include "vt.h"
10 #include "trace.h"
11 
12 static struct workqueue_struct *comp_vector_wq;
13 
14 /**
15  * rvt_cq_enter - add a new entry to the completion queue
16  * @cq: completion queue
17  * @entry: work completion entry to add
18  * @solicited: true if @entry is solicited
19  *
20  * This may be called with qp->s_lock held.
21  *
22  * Return: return true on success, else return
23  * false if cq is full.
24  */
rvt_cq_enter(struct rvt_cq * cq,struct ib_wc * entry,bool solicited)25 bool rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
26 {
27 	struct ib_uverbs_wc *uqueue = NULL;
28 	struct ib_wc *kqueue = NULL;
29 	struct rvt_cq_wc *u_wc = NULL;
30 	struct rvt_k_cq_wc *k_wc = NULL;
31 	unsigned long flags;
32 	u32 head;
33 	u32 next;
34 	u32 tail;
35 
36 	spin_lock_irqsave(&cq->lock, flags);
37 
38 	if (cq->ip) {
39 		u_wc = cq->queue;
40 		uqueue = &u_wc->uqueue[0];
41 		head = RDMA_READ_UAPI_ATOMIC(u_wc->head);
42 		tail = RDMA_READ_UAPI_ATOMIC(u_wc->tail);
43 	} else {
44 		k_wc = cq->kqueue;
45 		kqueue = &k_wc->kqueue[0];
46 		head = k_wc->head;
47 		tail = k_wc->tail;
48 	}
49 
50 	/*
51 	 * Note that the head pointer might be writable by
52 	 * user processes.Take care to verify it is a sane value.
53 	 */
54 	if (head >= (unsigned)cq->ibcq.cqe) {
55 		head = cq->ibcq.cqe;
56 		next = 0;
57 	} else {
58 		next = head + 1;
59 	}
60 
61 	if (unlikely(next == tail || cq->cq_full)) {
62 		struct rvt_dev_info *rdi = cq->rdi;
63 
64 		if (!cq->cq_full)
65 			rvt_pr_err_ratelimited(rdi, "CQ is full!\n");
66 		cq->cq_full = true;
67 		spin_unlock_irqrestore(&cq->lock, flags);
68 		if (cq->ibcq.event_handler) {
69 			struct ib_event ev;
70 
71 			ev.device = cq->ibcq.device;
72 			ev.element.cq = &cq->ibcq;
73 			ev.event = IB_EVENT_CQ_ERR;
74 			cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
75 		}
76 		return false;
77 	}
78 	trace_rvt_cq_enter(cq, entry, head);
79 	if (uqueue) {
80 		uqueue[head].wr_id = entry->wr_id;
81 		uqueue[head].status = entry->status;
82 		uqueue[head].opcode = entry->opcode;
83 		uqueue[head].vendor_err = entry->vendor_err;
84 		uqueue[head].byte_len = entry->byte_len;
85 		uqueue[head].ex.imm_data = entry->ex.imm_data;
86 		uqueue[head].qp_num = entry->qp->qp_num;
87 		uqueue[head].src_qp = entry->src_qp;
88 		uqueue[head].wc_flags = entry->wc_flags;
89 		uqueue[head].pkey_index = entry->pkey_index;
90 		uqueue[head].slid = ib_lid_cpu16(entry->slid);
91 		uqueue[head].sl = entry->sl;
92 		uqueue[head].dlid_path_bits = entry->dlid_path_bits;
93 		uqueue[head].port_num = entry->port_num;
94 		/* Make sure entry is written before the head index. */
95 		RDMA_WRITE_UAPI_ATOMIC(u_wc->head, next);
96 	} else {
97 		kqueue[head] = *entry;
98 		k_wc->head = next;
99 	}
100 
101 	if (cq->notify == IB_CQ_NEXT_COMP ||
102 	    (cq->notify == IB_CQ_SOLICITED &&
103 	     (solicited || entry->status != IB_WC_SUCCESS))) {
104 		/*
105 		 * This will cause send_complete() to be called in
106 		 * another thread.
107 		 */
108 		cq->notify = RVT_CQ_NONE;
109 		cq->triggered++;
110 		queue_work_on(cq->comp_vector_cpu, comp_vector_wq,
111 			      &cq->comptask);
112 	}
113 
114 	spin_unlock_irqrestore(&cq->lock, flags);
115 	return true;
116 }
117 EXPORT_SYMBOL(rvt_cq_enter);
118 
send_complete(struct work_struct * work)119 static void send_complete(struct work_struct *work)
120 {
121 	struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask);
122 
123 	/*
124 	 * The completion handler will most likely rearm the notification
125 	 * and poll for all pending entries.  If a new completion entry
126 	 * is added while we are in this routine, queue_work()
127 	 * won't call us again until we return so we check triggered to
128 	 * see if we need to call the handler again.
129 	 */
130 	for (;;) {
131 		u8 triggered = cq->triggered;
132 
133 		/*
134 		 * IPoIB connected mode assumes the callback is from a
135 		 * soft IRQ. We simulate this by blocking "bottom halves".
136 		 * See the implementation for ipoib_cm_handle_tx_wc(),
137 		 * netif_tx_lock_bh() and netif_tx_lock().
138 		 */
139 		local_bh_disable();
140 		cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
141 		local_bh_enable();
142 
143 		if (cq->triggered == triggered)
144 			return;
145 	}
146 }
147 
148 /**
149  * rvt_create_cq - create a completion queue
150  * @ibcq: Allocated CQ
151  * @attr: creation attributes
152  * @udata: user data for libibverbs.so
153  *
154  * Called by ib_create_cq() in the generic verbs code.
155  *
156  * Return: 0 on success
157  */
rvt_create_cq(struct ib_cq * ibcq,const struct ib_cq_init_attr * attr,struct ib_udata * udata)158 int rvt_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
159 		  struct ib_udata *udata)
160 {
161 	struct ib_device *ibdev = ibcq->device;
162 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
163 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
164 	struct rvt_cq_wc *u_wc = NULL;
165 	struct rvt_k_cq_wc *k_wc = NULL;
166 	u32 sz;
167 	unsigned int entries = attr->cqe;
168 	int comp_vector = attr->comp_vector;
169 	int err;
170 
171 	if (attr->flags)
172 		return -EOPNOTSUPP;
173 
174 	if (entries < 1 || entries > rdi->dparms.props.max_cqe)
175 		return -EINVAL;
176 
177 	if (comp_vector < 0)
178 		comp_vector = 0;
179 
180 	comp_vector = comp_vector % rdi->ibdev.num_comp_vectors;
181 
182 	/*
183 	 * Allocate the completion queue entries and head/tail pointers.
184 	 * This is allocated separately so that it can be resized and
185 	 * also mapped into user space.
186 	 * We need to use vmalloc() in order to support mmap and large
187 	 * numbers of entries.
188 	 */
189 	if (udata && udata->outlen >= sizeof(__u64)) {
190 		sz = sizeof(struct ib_uverbs_wc) * (entries + 1);
191 		sz += sizeof(*u_wc);
192 		u_wc = vmalloc_user(sz);
193 		if (!u_wc)
194 			return -ENOMEM;
195 	} else {
196 		sz = sizeof(struct ib_wc) * (entries + 1);
197 		sz += sizeof(*k_wc);
198 		k_wc = vzalloc_node(sz, rdi->dparms.node);
199 		if (!k_wc)
200 			return -ENOMEM;
201 	}
202 
203 	/*
204 	 * Return the address of the WC as the offset to mmap.
205 	 * See rvt_mmap() for details.
206 	 */
207 	if (udata && udata->outlen >= sizeof(__u64)) {
208 		cq->ip = rvt_create_mmap_info(rdi, sz, udata, u_wc);
209 		if (IS_ERR(cq->ip)) {
210 			err = PTR_ERR(cq->ip);
211 			goto bail_wc;
212 		}
213 
214 		err = ib_copy_to_udata(udata, &cq->ip->offset,
215 				       sizeof(cq->ip->offset));
216 		if (err)
217 			goto bail_ip;
218 	}
219 
220 	spin_lock_irq(&rdi->n_cqs_lock);
221 	if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) {
222 		spin_unlock_irq(&rdi->n_cqs_lock);
223 		err = -ENOMEM;
224 		goto bail_ip;
225 	}
226 
227 	rdi->n_cqs_allocated++;
228 	spin_unlock_irq(&rdi->n_cqs_lock);
229 
230 	if (cq->ip) {
231 		spin_lock_irq(&rdi->pending_lock);
232 		list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps);
233 		spin_unlock_irq(&rdi->pending_lock);
234 	}
235 
236 	/*
237 	 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
238 	 * The number of entries should be >= the number requested or return
239 	 * an error.
240 	 */
241 	cq->rdi = rdi;
242 	if (rdi->driver_f.comp_vect_cpu_lookup)
243 		cq->comp_vector_cpu =
244 			rdi->driver_f.comp_vect_cpu_lookup(rdi, comp_vector);
245 	else
246 		cq->comp_vector_cpu =
247 			cpumask_first(cpumask_of_node(rdi->dparms.node));
248 
249 	cq->ibcq.cqe = entries;
250 	cq->notify = RVT_CQ_NONE;
251 	spin_lock_init(&cq->lock);
252 	INIT_WORK(&cq->comptask, send_complete);
253 	if (u_wc)
254 		cq->queue = u_wc;
255 	else
256 		cq->kqueue = k_wc;
257 
258 	trace_rvt_create_cq(cq, attr);
259 	return 0;
260 
261 bail_ip:
262 	kfree(cq->ip);
263 bail_wc:
264 	vfree(u_wc);
265 	vfree(k_wc);
266 	return err;
267 }
268 
269 /**
270  * rvt_destroy_cq - destroy a completion queue
271  * @ibcq: the completion queue to destroy.
272  * @udata: user data or NULL for kernel object
273  *
274  * Called by ib_destroy_cq() in the generic verbs code.
275  */
rvt_destroy_cq(struct ib_cq * ibcq,struct ib_udata * udata)276 int rvt_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
277 {
278 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
279 	struct rvt_dev_info *rdi = cq->rdi;
280 
281 	flush_work(&cq->comptask);
282 	spin_lock_irq(&rdi->n_cqs_lock);
283 	rdi->n_cqs_allocated--;
284 	spin_unlock_irq(&rdi->n_cqs_lock);
285 	if (cq->ip)
286 		kref_put(&cq->ip->ref, rvt_release_mmap_info);
287 	else
288 		vfree(cq->kqueue);
289 	return 0;
290 }
291 
292 /**
293  * rvt_req_notify_cq - change the notification type for a completion queue
294  * @ibcq: the completion queue
295  * @notify_flags: the type of notification to request
296  *
297  * This may be called from interrupt context.  Also called by
298  * ib_req_notify_cq() in the generic verbs code.
299  *
300  * Return: 0 for success.
301  */
rvt_req_notify_cq(struct ib_cq * ibcq,enum ib_cq_notify_flags notify_flags)302 int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
303 {
304 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
305 	unsigned long flags;
306 	int ret = 0;
307 
308 	spin_lock_irqsave(&cq->lock, flags);
309 	/*
310 	 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
311 	 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
312 	 */
313 	if (cq->notify != IB_CQ_NEXT_COMP)
314 		cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
315 
316 	if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
317 		if (cq->queue) {
318 			if (RDMA_READ_UAPI_ATOMIC(cq->queue->head) !=
319 				RDMA_READ_UAPI_ATOMIC(cq->queue->tail))
320 				ret = 1;
321 		} else {
322 			if (cq->kqueue->head != cq->kqueue->tail)
323 				ret = 1;
324 		}
325 	}
326 
327 	spin_unlock_irqrestore(&cq->lock, flags);
328 
329 	return ret;
330 }
331 
332 /*
333  * rvt_resize_cq - change the size of the CQ
334  * @ibcq: the completion queue
335  *
336  * Return: 0 for success.
337  */
rvt_resize_cq(struct ib_cq * ibcq,int cqe,struct ib_udata * udata)338 int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
339 {
340 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
341 	u32 head, tail, n;
342 	int ret;
343 	u32 sz;
344 	struct rvt_dev_info *rdi = cq->rdi;
345 	struct rvt_cq_wc *u_wc = NULL;
346 	struct rvt_cq_wc *old_u_wc = NULL;
347 	struct rvt_k_cq_wc *k_wc = NULL;
348 	struct rvt_k_cq_wc *old_k_wc = NULL;
349 
350 	if (cqe < 1 || cqe > rdi->dparms.props.max_cqe)
351 		return -EINVAL;
352 
353 	/*
354 	 * Need to use vmalloc() if we want to support large #s of entries.
355 	 */
356 	if (udata && udata->outlen >= sizeof(__u64)) {
357 		sz = sizeof(struct ib_uverbs_wc) * (cqe + 1);
358 		sz += sizeof(*u_wc);
359 		u_wc = vmalloc_user(sz);
360 		if (!u_wc)
361 			return -ENOMEM;
362 	} else {
363 		sz = sizeof(struct ib_wc) * (cqe + 1);
364 		sz += sizeof(*k_wc);
365 		k_wc = vzalloc_node(sz, rdi->dparms.node);
366 		if (!k_wc)
367 			return -ENOMEM;
368 	}
369 	/* Check that we can write the offset to mmap. */
370 	if (udata && udata->outlen >= sizeof(__u64)) {
371 		__u64 offset = 0;
372 
373 		ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
374 		if (ret)
375 			goto bail_free;
376 	}
377 
378 	spin_lock_irq(&cq->lock);
379 	/*
380 	 * Make sure head and tail are sane since they
381 	 * might be user writable.
382 	 */
383 	if (u_wc) {
384 		old_u_wc = cq->queue;
385 		head = RDMA_READ_UAPI_ATOMIC(old_u_wc->head);
386 		tail = RDMA_READ_UAPI_ATOMIC(old_u_wc->tail);
387 	} else {
388 		old_k_wc = cq->kqueue;
389 		head = old_k_wc->head;
390 		tail = old_k_wc->tail;
391 	}
392 
393 	if (head > (u32)cq->ibcq.cqe)
394 		head = (u32)cq->ibcq.cqe;
395 	if (tail > (u32)cq->ibcq.cqe)
396 		tail = (u32)cq->ibcq.cqe;
397 	if (head < tail)
398 		n = cq->ibcq.cqe + 1 + head - tail;
399 	else
400 		n = head - tail;
401 	if (unlikely((u32)cqe < n)) {
402 		ret = -EINVAL;
403 		goto bail_unlock;
404 	}
405 	for (n = 0; tail != head; n++) {
406 		if (u_wc)
407 			u_wc->uqueue[n] = old_u_wc->uqueue[tail];
408 		else
409 			k_wc->kqueue[n] = old_k_wc->kqueue[tail];
410 		if (tail == (u32)cq->ibcq.cqe)
411 			tail = 0;
412 		else
413 			tail++;
414 	}
415 	cq->ibcq.cqe = cqe;
416 	if (u_wc) {
417 		RDMA_WRITE_UAPI_ATOMIC(u_wc->head, n);
418 		RDMA_WRITE_UAPI_ATOMIC(u_wc->tail, 0);
419 		cq->queue = u_wc;
420 	} else {
421 		k_wc->head = n;
422 		k_wc->tail = 0;
423 		cq->kqueue = k_wc;
424 	}
425 	spin_unlock_irq(&cq->lock);
426 
427 	if (u_wc)
428 		vfree(old_u_wc);
429 	else
430 		vfree(old_k_wc);
431 
432 	if (cq->ip) {
433 		struct rvt_mmap_info *ip = cq->ip;
434 
435 		rvt_update_mmap_info(rdi, ip, sz, u_wc);
436 
437 		/*
438 		 * Return the offset to mmap.
439 		 * See rvt_mmap() for details.
440 		 */
441 		if (udata && udata->outlen >= sizeof(__u64)) {
442 			ret = ib_copy_to_udata(udata, &ip->offset,
443 					       sizeof(ip->offset));
444 			if (ret)
445 				return ret;
446 		}
447 
448 		spin_lock_irq(&rdi->pending_lock);
449 		if (list_empty(&ip->pending_mmaps))
450 			list_add(&ip->pending_mmaps, &rdi->pending_mmaps);
451 		spin_unlock_irq(&rdi->pending_lock);
452 	}
453 
454 	return 0;
455 
456 bail_unlock:
457 	spin_unlock_irq(&cq->lock);
458 bail_free:
459 	vfree(u_wc);
460 	vfree(k_wc);
461 
462 	return ret;
463 }
464 
465 /**
466  * rvt_poll_cq - poll for work completion entries
467  * @ibcq: the completion queue to poll
468  * @num_entries: the maximum number of entries to return
469  * @entry: pointer to array where work completions are placed
470  *
471  * This may be called from interrupt context.  Also called by ib_poll_cq()
472  * in the generic verbs code.
473  *
474  * Return: the number of completion entries polled.
475  */
rvt_poll_cq(struct ib_cq * ibcq,int num_entries,struct ib_wc * entry)476 int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
477 {
478 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
479 	struct rvt_k_cq_wc *wc;
480 	unsigned long flags;
481 	int npolled;
482 	u32 tail;
483 
484 	/* The kernel can only poll a kernel completion queue */
485 	if (cq->ip)
486 		return -EINVAL;
487 
488 	spin_lock_irqsave(&cq->lock, flags);
489 
490 	wc = cq->kqueue;
491 	tail = wc->tail;
492 	if (tail > (u32)cq->ibcq.cqe)
493 		tail = (u32)cq->ibcq.cqe;
494 	for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
495 		if (tail == wc->head)
496 			break;
497 		/* The kernel doesn't need a RMB since it has the lock. */
498 		trace_rvt_cq_poll(cq, &wc->kqueue[tail], npolled);
499 		*entry = wc->kqueue[tail];
500 		if (tail >= cq->ibcq.cqe)
501 			tail = 0;
502 		else
503 			tail++;
504 	}
505 	wc->tail = tail;
506 
507 	spin_unlock_irqrestore(&cq->lock, flags);
508 
509 	return npolled;
510 }
511 
512 /**
513  * rvt_driver_cq_init - Init cq resources on behalf of driver
514  *
515  * Return: 0 on success
516  */
rvt_driver_cq_init(void)517 int rvt_driver_cq_init(void)
518 {
519 	comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE,
520 					 0, "rdmavt_cq");
521 	if (!comp_vector_wq)
522 		return -ENOMEM;
523 
524 	return 0;
525 }
526 
527 /**
528  * rvt_cq_exit - tear down cq reources
529  */
rvt_cq_exit(void)530 void rvt_cq_exit(void)
531 {
532 	destroy_workqueue(comp_vector_wq);
533 	comp_vector_wq = NULL;
534 }
535